1use anthropic::{AnthropicModelMode, parse_prompt_too_long};
2use anyhow::{Result, anyhow};
3use client::{Client, UserStore, zed_urls};
4use collections::BTreeMap;
5use feature_flags::{FeatureFlagAppExt, LlmClosedBetaFeatureFlag, ZedProFeatureFlag};
6use futures::{
7 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
8};
9use gpui::{AnyElement, AnyView, App, AsyncApp, Context, Entity, Subscription, Task};
10use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
11use language_model::{
12 AuthenticateError, CloudModel, CompletionRequestStatus, LanguageModel,
13 LanguageModelCacheConfiguration, LanguageModelCompletionError, LanguageModelId,
14 LanguageModelKnownError, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
15 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest,
16 LanguageModelToolSchemaFormat, ModelRequestLimitReachedError, RateLimiter, RequestUsage,
17 ZED_CLOUD_PROVIDER_ID,
18};
19use language_model::{
20 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken,
21 MaxMonthlySpendReachedError, PaymentRequiredError, RefreshLlmTokenListener,
22};
23use proto::Plan;
24use schemars::JsonSchema;
25use serde::{Deserialize, Serialize, de::DeserializeOwned};
26use settings::{Settings, SettingsStore};
27use smol::Timer;
28use smol::io::{AsyncReadExt, BufReader};
29use std::pin::Pin;
30use std::str::FromStr as _;
31use std::{
32 sync::{Arc, LazyLock},
33 time::Duration,
34};
35use strum::IntoEnumIterator;
36use thiserror::Error;
37use ui::{TintColor, prelude::*};
38use zed_llm_client::{
39 CURRENT_PLAN_HEADER_NAME, CompletionBody, CountTokensBody, CountTokensResponse,
40 EXPIRED_LLM_TOKEN_HEADER_NAME, MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
41 MODEL_REQUESTS_RESOURCE_HEADER_VALUE, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
42 TOOL_USE_LIMIT_REACHED_HEADER_NAME,
43};
44
45use crate::AllLanguageModelSettings;
46use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
47use crate::provider::google::{GoogleEventMapper, into_google};
48use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
49
50pub const PROVIDER_NAME: &str = "Zed";
51
52const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
53 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
54
55fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
56 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
57 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
58 .map(|json| serde_json::from_str(json).unwrap())
59 .unwrap_or_default()
60 });
61 ADDITIONAL_MODELS.as_slice()
62}
63
64#[derive(Default, Clone, Debug, PartialEq)]
65pub struct ZedDotDevSettings {
66 pub available_models: Vec<AvailableModel>,
67}
68
69#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
70#[serde(rename_all = "lowercase")]
71pub enum AvailableProvider {
72 Anthropic,
73 OpenAi,
74 Google,
75}
76
77#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
78pub struct AvailableModel {
79 /// The provider of the language model.
80 pub provider: AvailableProvider,
81 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
82 pub name: String,
83 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
84 pub display_name: Option<String>,
85 /// The size of the context window, indicating the maximum number of tokens the model can process.
86 pub max_tokens: usize,
87 /// The maximum number of output tokens allowed by the model.
88 pub max_output_tokens: Option<u32>,
89 /// The maximum number of completion tokens allowed by the model (o1-* only)
90 pub max_completion_tokens: Option<u32>,
91 /// Override this model with a different Anthropic model for tool calls.
92 pub tool_override: Option<String>,
93 /// Indicates whether this custom model supports caching.
94 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
95 /// The default temperature to use for this model.
96 pub default_temperature: Option<f32>,
97 /// Any extra beta headers to provide when using the model.
98 #[serde(default)]
99 pub extra_beta_headers: Vec<String>,
100 /// The model's mode (e.g. thinking)
101 pub mode: Option<ModelMode>,
102}
103
104#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
105#[serde(tag = "type", rename_all = "lowercase")]
106pub enum ModelMode {
107 #[default]
108 Default,
109 Thinking {
110 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
111 budget_tokens: Option<u32>,
112 },
113}
114
115impl From<ModelMode> for AnthropicModelMode {
116 fn from(value: ModelMode) -> Self {
117 match value {
118 ModelMode::Default => AnthropicModelMode::Default,
119 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
120 }
121 }
122}
123
124pub struct CloudLanguageModelProvider {
125 client: Arc<Client>,
126 state: gpui::Entity<State>,
127 _maintain_client_status: Task<()>,
128}
129
130pub struct State {
131 client: Arc<Client>,
132 llm_api_token: LlmApiToken,
133 user_store: Entity<UserStore>,
134 status: client::Status,
135 accept_terms: Option<Task<Result<()>>>,
136 _settings_subscription: Subscription,
137 _llm_token_subscription: Subscription,
138}
139
140impl State {
141 fn new(
142 client: Arc<Client>,
143 user_store: Entity<UserStore>,
144 status: client::Status,
145 cx: &mut Context<Self>,
146 ) -> Self {
147 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
148
149 Self {
150 client: client.clone(),
151 llm_api_token: LlmApiToken::default(),
152 user_store,
153 status,
154 accept_terms: None,
155 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
156 cx.notify();
157 }),
158 _llm_token_subscription: cx.subscribe(
159 &refresh_llm_token_listener,
160 |this, _listener, _event, cx| {
161 let client = this.client.clone();
162 let llm_api_token = this.llm_api_token.clone();
163 cx.spawn(async move |_this, _cx| {
164 llm_api_token.refresh(&client).await?;
165 anyhow::Ok(())
166 })
167 .detach_and_log_err(cx);
168 },
169 ),
170 }
171 }
172
173 fn is_signed_out(&self) -> bool {
174 self.status.is_signed_out()
175 }
176
177 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
178 let client = self.client.clone();
179 cx.spawn(async move |this, cx| {
180 client.authenticate_and_connect(true, &cx).await?;
181 this.update(cx, |_, cx| cx.notify())
182 })
183 }
184
185 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
186 self.user_store
187 .read(cx)
188 .current_user_has_accepted_terms()
189 .unwrap_or(false)
190 }
191
192 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
193 let user_store = self.user_store.clone();
194 self.accept_terms = Some(cx.spawn(async move |this, cx| {
195 let _ = user_store
196 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
197 .await;
198 this.update(cx, |this, cx| {
199 this.accept_terms = None;
200 cx.notify()
201 })
202 }));
203 }
204}
205
206impl CloudLanguageModelProvider {
207 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
208 let mut status_rx = client.status();
209 let status = *status_rx.borrow();
210
211 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
212
213 let state_ref = state.downgrade();
214 let maintain_client_status = cx.spawn(async move |cx| {
215 while let Some(status) = status_rx.next().await {
216 if let Some(this) = state_ref.upgrade() {
217 _ = this.update(cx, |this, cx| {
218 if this.status != status {
219 this.status = status;
220 cx.notify();
221 }
222 });
223 } else {
224 break;
225 }
226 }
227 });
228
229 Self {
230 client,
231 state: state.clone(),
232 _maintain_client_status: maintain_client_status,
233 }
234 }
235
236 fn create_language_model(
237 &self,
238 model: CloudModel,
239 llm_api_token: LlmApiToken,
240 ) -> Arc<dyn LanguageModel> {
241 Arc::new(CloudLanguageModel {
242 id: LanguageModelId::from(model.id().to_string()),
243 model,
244 llm_api_token: llm_api_token.clone(),
245 client: self.client.clone(),
246 request_limiter: RateLimiter::new(4),
247 })
248 }
249}
250
251impl LanguageModelProviderState for CloudLanguageModelProvider {
252 type ObservableEntity = State;
253
254 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
255 Some(self.state.clone())
256 }
257}
258
259impl LanguageModelProvider for CloudLanguageModelProvider {
260 fn id(&self) -> LanguageModelProviderId {
261 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
262 }
263
264 fn name(&self) -> LanguageModelProviderName {
265 LanguageModelProviderName(PROVIDER_NAME.into())
266 }
267
268 fn icon(&self) -> IconName {
269 IconName::AiZed
270 }
271
272 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
273 let llm_api_token = self.state.read(cx).llm_api_token.clone();
274 let model = CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet);
275 Some(self.create_language_model(model, llm_api_token))
276 }
277
278 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
279 let llm_api_token = self.state.read(cx).llm_api_token.clone();
280 let model = CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet);
281 Some(self.create_language_model(model, llm_api_token))
282 }
283
284 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
285 let llm_api_token = self.state.read(cx).llm_api_token.clone();
286 [
287 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
288 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
289 ]
290 .into_iter()
291 .map(|model| self.create_language_model(model, llm_api_token.clone()))
292 .collect()
293 }
294
295 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
296 let mut models = BTreeMap::default();
297
298 if cx.is_staff() {
299 for model in anthropic::Model::iter() {
300 if !matches!(model, anthropic::Model::Custom { .. }) {
301 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
302 }
303 }
304 for model in open_ai::Model::iter() {
305 if !matches!(model, open_ai::Model::Custom { .. }) {
306 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
307 }
308 }
309 for model in google_ai::Model::iter() {
310 if !matches!(model, google_ai::Model::Custom { .. }) {
311 models.insert(model.id().to_string(), CloudModel::Google(model));
312 }
313 }
314 } else {
315 models.insert(
316 anthropic::Model::Claude3_5Sonnet.id().to_string(),
317 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
318 );
319 models.insert(
320 anthropic::Model::Claude3_7Sonnet.id().to_string(),
321 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
322 );
323 models.insert(
324 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
325 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
326 );
327 }
328
329 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBetaFeatureFlag>() {
330 zed_cloud_provider_additional_models()
331 } else {
332 &[]
333 };
334
335 // Override with available models from settings
336 for model in AllLanguageModelSettings::get_global(cx)
337 .zed_dot_dev
338 .available_models
339 .iter()
340 .chain(llm_closed_beta_models)
341 .cloned()
342 {
343 let model = match model.provider {
344 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
345 name: model.name.clone(),
346 display_name: model.display_name.clone(),
347 max_tokens: model.max_tokens,
348 tool_override: model.tool_override.clone(),
349 cache_configuration: model.cache_configuration.as_ref().map(|config| {
350 anthropic::AnthropicModelCacheConfiguration {
351 max_cache_anchors: config.max_cache_anchors,
352 should_speculate: config.should_speculate,
353 min_total_token: config.min_total_token,
354 }
355 }),
356 default_temperature: model.default_temperature,
357 max_output_tokens: model.max_output_tokens,
358 extra_beta_headers: model.extra_beta_headers.clone(),
359 mode: model.mode.unwrap_or_default().into(),
360 }),
361 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
362 name: model.name.clone(),
363 display_name: model.display_name.clone(),
364 max_tokens: model.max_tokens,
365 max_output_tokens: model.max_output_tokens,
366 max_completion_tokens: model.max_completion_tokens,
367 }),
368 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
369 name: model.name.clone(),
370 display_name: model.display_name.clone(),
371 max_tokens: model.max_tokens,
372 }),
373 };
374 models.insert(model.id().to_string(), model.clone());
375 }
376
377 let llm_api_token = self.state.read(cx).llm_api_token.clone();
378 models
379 .into_values()
380 .map(|model| self.create_language_model(model, llm_api_token.clone()))
381 .collect()
382 }
383
384 fn is_authenticated(&self, cx: &App) -> bool {
385 !self.state.read(cx).is_signed_out()
386 }
387
388 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
389 Task::ready(Ok(()))
390 }
391
392 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
393 cx.new(|_| ConfigurationView {
394 state: self.state.clone(),
395 })
396 .into()
397 }
398
399 fn must_accept_terms(&self, cx: &App) -> bool {
400 !self.state.read(cx).has_accepted_terms_of_service(cx)
401 }
402
403 fn render_accept_terms(
404 &self,
405 view: LanguageModelProviderTosView,
406 cx: &mut App,
407 ) -> Option<AnyElement> {
408 render_accept_terms(self.state.clone(), view, cx)
409 }
410
411 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
412 Task::ready(Ok(()))
413 }
414}
415
416fn render_accept_terms(
417 state: Entity<State>,
418 view_kind: LanguageModelProviderTosView,
419 cx: &mut App,
420) -> Option<AnyElement> {
421 if state.read(cx).has_accepted_terms_of_service(cx) {
422 return None;
423 }
424
425 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
426
427 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
428 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadtEmptyState);
429
430 let terms_button = Button::new("terms_of_service", "Terms of Service")
431 .style(ButtonStyle::Subtle)
432 .icon(IconName::ArrowUpRight)
433 .icon_color(Color::Muted)
434 .icon_size(IconSize::XSmall)
435 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
436 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
437
438 let button_container = h_flex().child(
439 Button::new("accept_terms", "I accept the Terms of Service")
440 .when(!thread_empty_state, |this| {
441 this.full_width()
442 .style(ButtonStyle::Tinted(TintColor::Accent))
443 .icon(IconName::Check)
444 .icon_position(IconPosition::Start)
445 .icon_size(IconSize::Small)
446 })
447 .when(thread_empty_state, |this| {
448 this.style(ButtonStyle::Tinted(TintColor::Warning))
449 .label_size(LabelSize::Small)
450 })
451 .disabled(accept_terms_disabled)
452 .on_click({
453 let state = state.downgrade();
454 move |_, _window, cx| {
455 state
456 .update(cx, |state, cx| state.accept_terms_of_service(cx))
457 .ok();
458 }
459 }),
460 );
461
462 let form = if thread_empty_state {
463 h_flex()
464 .w_full()
465 .flex_wrap()
466 .justify_between()
467 .child(
468 h_flex()
469 .child(
470 Label::new("To start using Zed AI, please read and accept the")
471 .size(LabelSize::Small),
472 )
473 .child(terms_button),
474 )
475 .child(button_container)
476 } else {
477 v_flex()
478 .w_full()
479 .gap_2()
480 .child(
481 h_flex()
482 .flex_wrap()
483 .when(thread_fresh_start, |this| this.justify_center())
484 .child(Label::new(
485 "To start using Zed AI, please read and accept the",
486 ))
487 .child(terms_button),
488 )
489 .child({
490 match view_kind {
491 LanguageModelProviderTosView::PromptEditorPopup => {
492 button_container.w_full().justify_end()
493 }
494 LanguageModelProviderTosView::Configuration => {
495 button_container.w_full().justify_start()
496 }
497 LanguageModelProviderTosView::ThreadFreshStart => {
498 button_container.w_full().justify_center()
499 }
500 LanguageModelProviderTosView::ThreadtEmptyState => div().w_0(),
501 }
502 })
503 };
504
505 Some(form.into_any())
506}
507
508pub struct CloudLanguageModel {
509 id: LanguageModelId,
510 model: CloudModel,
511 llm_api_token: LlmApiToken,
512 client: Arc<Client>,
513 request_limiter: RateLimiter,
514}
515
516struct PerformLlmCompletionResponse {
517 response: Response<AsyncBody>,
518 usage: Option<RequestUsage>,
519 tool_use_limit_reached: bool,
520 includes_queue_events: bool,
521}
522
523impl CloudLanguageModel {
524 const MAX_RETRIES: usize = 3;
525
526 async fn perform_llm_completion(
527 client: Arc<Client>,
528 llm_api_token: LlmApiToken,
529 body: CompletionBody,
530 ) -> Result<PerformLlmCompletionResponse> {
531 let http_client = &client.http_client();
532
533 let mut token = llm_api_token.acquire(&client).await?;
534 let mut retries_remaining = Self::MAX_RETRIES;
535 let mut retry_delay = Duration::from_secs(1);
536
537 loop {
538 let request_builder = http_client::Request::builder().method(Method::POST);
539 let request_builder = if let Ok(completions_url) = std::env::var("ZED_COMPLETIONS_URL")
540 {
541 request_builder.uri(completions_url)
542 } else {
543 request_builder.uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
544 };
545 let request = request_builder
546 .header("Content-Type", "application/json")
547 .header("Authorization", format!("Bearer {token}"))
548 .header("x-zed-client-supports-queueing", "true")
549 .body(serde_json::to_string(&body)?.into())?;
550 let mut response = http_client.send(request).await?;
551 let status = response.status();
552 if status.is_success() {
553 let includes_queue_events = response
554 .headers()
555 .get("x-zed-server-supports-queueing")
556 .is_some();
557 let tool_use_limit_reached = response
558 .headers()
559 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
560 .is_some();
561 let usage = RequestUsage::from_headers(response.headers()).ok();
562
563 return Ok(PerformLlmCompletionResponse {
564 response,
565 usage,
566 includes_queue_events,
567 tool_use_limit_reached,
568 });
569 } else if response
570 .headers()
571 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
572 .is_some()
573 {
574 retries_remaining -= 1;
575 token = llm_api_token.refresh(&client).await?;
576 } else if status == StatusCode::FORBIDDEN
577 && response
578 .headers()
579 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
580 .is_some()
581 {
582 return Err(anyhow!(MaxMonthlySpendReachedError));
583 } else if status == StatusCode::FORBIDDEN
584 && response
585 .headers()
586 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
587 .is_some()
588 {
589 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
590 .headers()
591 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
592 .and_then(|resource| resource.to_str().ok())
593 {
594 if let Some(plan) = response
595 .headers()
596 .get(CURRENT_PLAN_HEADER_NAME)
597 .and_then(|plan| plan.to_str().ok())
598 .and_then(|plan| zed_llm_client::Plan::from_str(plan).ok())
599 {
600 let plan = match plan {
601 zed_llm_client::Plan::Free => Plan::Free,
602 zed_llm_client::Plan::ZedPro => Plan::ZedPro,
603 zed_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
604 };
605 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
606 }
607 }
608
609 return Err(anyhow!("Forbidden"));
610 } else if status.as_u16() >= 500 && status.as_u16() < 600 {
611 // If we encounter an error in the 500 range, retry after a delay.
612 // We've seen at least these in the wild from API providers:
613 // * 500 Internal Server Error
614 // * 502 Bad Gateway
615 // * 529 Service Overloaded
616
617 if retries_remaining == 0 {
618 let mut body = String::new();
619 response.body_mut().read_to_string(&mut body).await?;
620 return Err(anyhow!(
621 "cloud language model completion failed after {} retries with status {status}: {body}",
622 Self::MAX_RETRIES
623 ));
624 }
625
626 Timer::after(retry_delay).await;
627
628 retries_remaining -= 1;
629 retry_delay *= 2; // If it fails again, wait longer.
630 } else if status == StatusCode::PAYMENT_REQUIRED {
631 return Err(anyhow!(PaymentRequiredError));
632 } else {
633 let mut body = String::new();
634 response.body_mut().read_to_string(&mut body).await?;
635 return Err(anyhow!(ApiError { status, body }));
636 }
637 }
638 }
639}
640
641#[derive(Debug, Error)]
642#[error("cloud language model completion failed with status {status}: {body}")]
643struct ApiError {
644 status: StatusCode,
645 body: String,
646}
647
648impl LanguageModel for CloudLanguageModel {
649 fn id(&self) -> LanguageModelId {
650 self.id.clone()
651 }
652
653 fn name(&self) -> LanguageModelName {
654 LanguageModelName::from(self.model.display_name().to_string())
655 }
656
657 fn provider_id(&self) -> LanguageModelProviderId {
658 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
659 }
660
661 fn provider_name(&self) -> LanguageModelProviderName {
662 LanguageModelProviderName(PROVIDER_NAME.into())
663 }
664
665 fn supports_tools(&self) -> bool {
666 match self.model {
667 CloudModel::Anthropic(_) => true,
668 CloudModel::Google(_) => true,
669 CloudModel::OpenAi(_) => true,
670 }
671 }
672
673 fn telemetry_id(&self) -> String {
674 format!("zed.dev/{}", self.model.id())
675 }
676
677 fn availability(&self) -> LanguageModelAvailability {
678 self.model.availability()
679 }
680
681 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
682 self.model.tool_input_format()
683 }
684
685 fn max_token_count(&self) -> usize {
686 self.model.max_token_count()
687 }
688
689 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
690 match &self.model {
691 CloudModel::Anthropic(model) => {
692 model
693 .cache_configuration()
694 .map(|cache| LanguageModelCacheConfiguration {
695 max_cache_anchors: cache.max_cache_anchors,
696 should_speculate: cache.should_speculate,
697 min_total_token: cache.min_total_token,
698 })
699 }
700 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
701 }
702 }
703
704 fn count_tokens(
705 &self,
706 request: LanguageModelRequest,
707 cx: &App,
708 ) -> BoxFuture<'static, Result<usize>> {
709 match self.model.clone() {
710 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
711 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
712 CloudModel::Google(model) => {
713 let client = self.client.clone();
714 let llm_api_token = self.llm_api_token.clone();
715 let request = into_google(request, model.id().into());
716 async move {
717 let http_client = &client.http_client();
718 let token = llm_api_token.acquire(&client).await?;
719
720 let request_builder = http_client::Request::builder().method(Method::POST);
721 let request_builder =
722 if let Ok(completions_url) = std::env::var("ZED_COUNT_TOKENS_URL") {
723 request_builder.uri(completions_url)
724 } else {
725 request_builder.uri(
726 http_client
727 .build_zed_llm_url("/count_tokens", &[])?
728 .as_ref(),
729 )
730 };
731 let request_body = CountTokensBody {
732 provider: zed_llm_client::LanguageModelProvider::Google,
733 model: model.id().into(),
734 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
735 contents: request.contents,
736 })?,
737 };
738 let request = request_builder
739 .header("Content-Type", "application/json")
740 .header("Authorization", format!("Bearer {token}"))
741 .body(serde_json::to_string(&request_body)?.into())?;
742 let mut response = http_client.send(request).await?;
743 let status = response.status();
744 let mut response_body = String::new();
745 response
746 .body_mut()
747 .read_to_string(&mut response_body)
748 .await?;
749
750 if status.is_success() {
751 let response_body: CountTokensResponse =
752 serde_json::from_str(&response_body)?;
753
754 Ok(response_body.tokens)
755 } else {
756 Err(anyhow!(ApiError {
757 status,
758 body: response_body
759 }))
760 }
761 }
762 .boxed()
763 }
764 }
765 }
766
767 fn stream_completion(
768 &self,
769 request: LanguageModelRequest,
770 cx: &AsyncApp,
771 ) -> BoxFuture<
772 'static,
773 Result<
774 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
775 >,
776 > {
777 self.stream_completion_with_usage(request, cx)
778 .map(|result| result.map(|(stream, _)| stream))
779 .boxed()
780 }
781
782 fn stream_completion_with_usage(
783 &self,
784 request: LanguageModelRequest,
785 _cx: &AsyncApp,
786 ) -> BoxFuture<
787 'static,
788 Result<(
789 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
790 Option<RequestUsage>,
791 )>,
792 > {
793 let thread_id = request.thread_id.clone();
794 let prompt_id = request.prompt_id.clone();
795 let mode = request.mode;
796 match &self.model {
797 CloudModel::Anthropic(model) => {
798 let request = into_anthropic(
799 request,
800 model.request_id().into(),
801 model.default_temperature(),
802 model.max_output_tokens(),
803 model.mode(),
804 );
805 let client = self.client.clone();
806 let llm_api_token = self.llm_api_token.clone();
807 let future = self.request_limiter.stream_with_usage(async move {
808 let PerformLlmCompletionResponse {
809 response,
810 usage,
811 includes_queue_events,
812 tool_use_limit_reached,
813 } = Self::perform_llm_completion(
814 client.clone(),
815 llm_api_token,
816 CompletionBody {
817 thread_id,
818 prompt_id,
819 mode,
820 provider: zed_llm_client::LanguageModelProvider::Anthropic,
821 model: request.model.clone(),
822 provider_request: serde_json::to_value(&request)?,
823 },
824 )
825 .await
826 .map_err(|err| match err.downcast::<ApiError>() {
827 Ok(api_err) => {
828 if api_err.status == StatusCode::BAD_REQUEST {
829 if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
830 return anyhow!(
831 LanguageModelKnownError::ContextWindowLimitExceeded {
832 tokens
833 }
834 );
835 }
836 }
837 anyhow!(api_err)
838 }
839 Err(err) => anyhow!(err),
840 })?;
841
842 let mut mapper = AnthropicEventMapper::new();
843 Ok((
844 map_cloud_completion_events(
845 Box::pin(
846 response_lines(response, includes_queue_events)
847 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
848 ),
849 move |event| mapper.map_event(event),
850 ),
851 usage,
852 ))
853 });
854 async move {
855 let (stream, usage) = future.await?;
856 Ok((stream.boxed(), usage))
857 }
858 .boxed()
859 }
860 CloudModel::OpenAi(model) => {
861 let client = self.client.clone();
862 let request = into_open_ai(request, model, model.max_output_tokens());
863 let llm_api_token = self.llm_api_token.clone();
864 let future = self.request_limiter.stream_with_usage(async move {
865 let PerformLlmCompletionResponse {
866 response,
867 usage,
868 includes_queue_events,
869 tool_use_limit_reached,
870 } = Self::perform_llm_completion(
871 client.clone(),
872 llm_api_token,
873 CompletionBody {
874 thread_id,
875 prompt_id,
876 mode,
877 provider: zed_llm_client::LanguageModelProvider::OpenAi,
878 model: request.model.clone(),
879 provider_request: serde_json::to_value(&request)?,
880 },
881 )
882 .await?;
883
884 let mut mapper = OpenAiEventMapper::new();
885 Ok((
886 map_cloud_completion_events(
887 Box::pin(
888 response_lines(response, includes_queue_events)
889 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
890 ),
891 move |event| mapper.map_event(event),
892 ),
893 usage,
894 ))
895 });
896 async move {
897 let (stream, usage) = future.await?;
898 Ok((stream.boxed(), usage))
899 }
900 .boxed()
901 }
902 CloudModel::Google(model) => {
903 let client = self.client.clone();
904 let request = into_google(request, model.id().into());
905 let llm_api_token = self.llm_api_token.clone();
906 let future = self.request_limiter.stream_with_usage(async move {
907 let PerformLlmCompletionResponse {
908 response,
909 usage,
910 includes_queue_events,
911 tool_use_limit_reached,
912 } = Self::perform_llm_completion(
913 client.clone(),
914 llm_api_token,
915 CompletionBody {
916 thread_id,
917 prompt_id,
918 mode,
919 provider: zed_llm_client::LanguageModelProvider::Google,
920 model: request.model.clone(),
921 provider_request: serde_json::to_value(&request)?,
922 },
923 )
924 .await?;
925
926 let mut mapper = GoogleEventMapper::new();
927 Ok((
928 map_cloud_completion_events(
929 Box::pin(
930 response_lines(response, includes_queue_events)
931 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
932 ),
933 move |event| mapper.map_event(event),
934 ),
935 usage,
936 ))
937 });
938 async move {
939 let (stream, usage) = future.await?;
940 Ok((stream.boxed(), usage))
941 }
942 .boxed()
943 }
944 }
945 }
946}
947
948#[derive(Serialize, Deserialize)]
949#[serde(rename_all = "snake_case")]
950pub enum CloudCompletionEvent<T> {
951 System(CompletionRequestStatus),
952 Event(T),
953}
954
955fn map_cloud_completion_events<T, F>(
956 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
957 mut map_callback: F,
958) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
959where
960 T: DeserializeOwned + 'static,
961 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
962 + Send
963 + 'static,
964{
965 stream
966 .flat_map(move |event| {
967 futures::stream::iter(match event {
968 Err(error) => {
969 vec![Err(LanguageModelCompletionError::Other(error))]
970 }
971 Ok(CloudCompletionEvent::System(event)) => {
972 vec![Ok(LanguageModelCompletionEvent::QueueUpdate(event))]
973 }
974 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
975 })
976 })
977 .boxed()
978}
979
980fn tool_use_limit_reached_event<T>(
981 tool_use_limit_reached: bool,
982) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
983 futures::stream::iter(tool_use_limit_reached.then(|| {
984 Ok(CloudCompletionEvent::System(
985 CompletionRequestStatus::ToolUseLimitReached,
986 ))
987 }))
988}
989
990fn response_lines<T: DeserializeOwned>(
991 response: Response<AsyncBody>,
992 includes_queue_events: bool,
993) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
994 futures::stream::try_unfold(
995 (String::new(), BufReader::new(response.into_body())),
996 move |(mut line, mut body)| async move {
997 match body.read_line(&mut line).await {
998 Ok(0) => Ok(None),
999 Ok(_) => {
1000 let event = if includes_queue_events {
1001 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1002 } else {
1003 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1004 };
1005
1006 line.clear();
1007 Ok(Some((event, (line, body))))
1008 }
1009 Err(e) => Err(e.into()),
1010 }
1011 },
1012 )
1013}
1014
1015struct ConfigurationView {
1016 state: gpui::Entity<State>,
1017}
1018
1019impl ConfigurationView {
1020 fn authenticate(&mut self, cx: &mut Context<Self>) {
1021 self.state.update(cx, |state, cx| {
1022 state.authenticate(cx).detach_and_log_err(cx);
1023 });
1024 cx.notify();
1025 }
1026}
1027
1028impl Render for ConfigurationView {
1029 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1030 const ZED_AI_URL: &str = "https://zed.dev/ai";
1031
1032 let is_connected = !self.state.read(cx).is_signed_out();
1033 let plan = self.state.read(cx).user_store.read(cx).current_plan();
1034 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
1035
1036 let is_pro = plan == Some(proto::Plan::ZedPro);
1037 let subscription_text = Label::new(if is_pro {
1038 "You have full access to Zed's hosted LLMs, which include models from Anthropic, OpenAI, and Google. They come with faster speeds and higher limits through Zed Pro."
1039 } else {
1040 "You have basic access to models from Anthropic through the Zed AI Free plan."
1041 });
1042 let manage_subscription_button = if is_pro {
1043 Some(
1044 h_flex().child(
1045 Button::new("manage_settings", "Manage Subscription")
1046 .style(ButtonStyle::Tinted(TintColor::Accent))
1047 .on_click(
1048 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1049 ),
1050 ),
1051 )
1052 } else if cx.has_flag::<ZedProFeatureFlag>() {
1053 Some(
1054 h_flex()
1055 .gap_2()
1056 .child(
1057 Button::new("learn_more", "Learn more")
1058 .style(ButtonStyle::Subtle)
1059 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
1060 )
1061 .child(
1062 Button::new("upgrade", "Upgrade")
1063 .style(ButtonStyle::Subtle)
1064 .color(Color::Accent)
1065 .on_click(
1066 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1067 ),
1068 ),
1069 )
1070 } else {
1071 None
1072 };
1073
1074 if is_connected {
1075 v_flex()
1076 .gap_3()
1077 .w_full()
1078 .children(render_accept_terms(
1079 self.state.clone(),
1080 LanguageModelProviderTosView::Configuration,
1081 cx,
1082 ))
1083 .when(has_accepted_terms, |this| {
1084 this.child(subscription_text)
1085 .children(manage_subscription_button)
1086 })
1087 } else {
1088 v_flex()
1089 .gap_2()
1090 .child(Label::new("Use Zed AI to access hosted language models."))
1091 .child(
1092 Button::new("sign_in", "Sign In")
1093 .icon_color(Color::Muted)
1094 .icon(IconName::Github)
1095 .icon_position(IconPosition::Start)
1096 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
1097 )
1098 }
1099 }
1100}